From Oracle to Noisy Context: Mitigating Contextual Exposure Bias in Speech-LLMs

1. 1Use teacher error knowledge by incorporating error-prone history into training data
2. 2Implement a unified training framework to combine oracle and noisy context
3. 3Evaluate the model's performance under various noise conditions to ensure robustness
4. 4Fine-tune the model with noisy context to adapt to real-world scenarios

ML researchers and engineers working on speech-LLMs can benefit from this framework to improve the robustness of their models under realistic conversation histories

💡 Using a unified training framework with teacher error knowledge and noisy context can improve the robustness of speech-LLMs under realistic conversation histories

🗣️ Mitigate contextual exposure bias in speech-LLMs with teacher error knowledge and noisy context! 💡